regnault.dev
forked from
tangled.org/core
Monorepo for Tangled
boltless.me
1package nixery
2
3import (
4 "context"
5 "errors"
6 "fmt"
7 "io"
8 "log/slog"
9 "os"
10 "path"
11 "runtime"
12 "sync"
13 "time"
14
15 "github.com/docker/docker/api/types/container"
16 "github.com/docker/docker/api/types/image"
17 "github.com/docker/docker/api/types/mount"
18 "github.com/docker/docker/api/types/network"
19 "github.com/docker/docker/client"
20 "github.com/docker/docker/pkg/stdcopy"
21 "gopkg.in/yaml.v3"
22 "tangled.org/core/api/tangled"
23 "tangled.org/core/log"
24 "tangled.org/core/spindle/config"
25 "tangled.org/core/spindle/engine"
26 "tangled.org/core/spindle/models"
27 "tangled.org/core/spindle/secrets"
28)
29
30const (
31 workspaceDir = "/tangled/workspace"
32 homeDir = "/tangled/home"
33)
34
35type cleanupFunc func(context.Context) error
36
37type Engine struct {
38 docker client.APIClient
39 l *slog.Logger
40 cfg *config.Config
41
42 cleanupMu sync.Mutex
43 cleanup map[string][]cleanupFunc
44}
45
46type Step struct {
47 name string
48 kind models.StepKind
49 command string
50 environment map[string]string
51}
52
53func (s Step) Name() string {
54 return s.name
55}
56
57func (s Step) Command() string {
58 return s.command
59}
60
61func (s Step) Kind() models.StepKind {
62 return s.kind
63}
64
65// setupSteps get added to start of Steps
66type setupSteps []models.Step
67
68// addStep adds a step to the beginning of the workflow's steps.
69func (ss *setupSteps) addStep(step models.Step) {
70 *ss = append(*ss, step)
71}
72
73type addlFields struct {
74 image string
75 container string
76}
77
78func (e *Engine) InitWorkflow(twf tangled.Pipeline_Workflow, tpl tangled.Pipeline) (*models.Workflow, error) {
79 swf := &models.Workflow{}
80 addl := addlFields{}
81
82 dwf := &struct {
83 Steps []struct {
84 Command string `yaml:"command"`
85 Name string `yaml:"name"`
86 Environment map[string]string `yaml:"environment"`
87 } `yaml:"steps"`
88 Dependencies map[string][]string `yaml:"dependencies"`
89 Environment map[string]string `yaml:"environment"`
90 }{}
91 err := yaml.Unmarshal([]byte(twf.Raw), &dwf)
92 if err != nil {
93 return nil, err
94 }
95
96 for _, dstep := range dwf.Steps {
97 sstep := Step{}
98 sstep.environment = dstep.Environment
99 sstep.command = dstep.Command
100 sstep.name = dstep.Name
101 sstep.kind = models.StepKindUser
102 swf.Steps = append(swf.Steps, sstep)
103 }
104 swf.Name = twf.Name
105 swf.Environment = dwf.Environment
106 addl.image = workflowImage(dwf.Dependencies, e.cfg.NixeryPipelines.Nixery)
107
108 setup := &setupSteps{}
109
110 setup.addStep(nixConfStep())
111 setup.addStep(models.BuildCloneStep(twf, *tpl.TriggerMetadata, e.cfg.Server.Dev))
112 // this step could be empty
113 if s := dependencyStep(dwf.Dependencies); s != nil {
114 setup.addStep(*s)
115 }
116
117 // append setup steps in order to the start of workflow steps
118 swf.Steps = append(*setup, swf.Steps...)
119 swf.Data = addl
120
121 return swf, nil
122}
123
124func (e *Engine) WorkflowTimeout() time.Duration {
125 workflowTimeoutStr := e.cfg.NixeryPipelines.WorkflowTimeout
126 workflowTimeout, err := time.ParseDuration(workflowTimeoutStr)
127 if err != nil {
128 e.l.Error("failed to parse workflow timeout", "error", err, "timeout", workflowTimeoutStr)
129 workflowTimeout = 5 * time.Minute
130 }
131
132 return workflowTimeout
133}
134
135func workflowImage(deps map[string][]string, nixery string) string {
136 var dependencies string
137 for reg, ds := range deps {
138 if reg == "nixpkgs" {
139 dependencies = path.Join(ds...)
140 }
141 }
142
143 // load defaults from somewhere else
144 dependencies = path.Join(dependencies, "bash", "git", "coreutils", "nix")
145
146 if runtime.GOARCH == "arm64" {
147 dependencies = path.Join("arm64", dependencies)
148 }
149
150 return path.Join(nixery, dependencies)
151}
152
153func New(ctx context.Context, cfg *config.Config) (*Engine, error) {
154 dcli, err := client.NewClientWithOpts(client.FromEnv, client.WithAPIVersionNegotiation())
155 if err != nil {
156 return nil, err
157 }
158
159 l := log.FromContext(ctx).With("component", "spindle")
160
161 e := &Engine{
162 docker: dcli,
163 l: l,
164 cfg: cfg,
165 }
166
167 e.cleanup = make(map[string][]cleanupFunc)
168
169 return e, nil
170}
171
172func (e *Engine) SetupWorkflow(ctx context.Context, wid models.WorkflowId, wf *models.Workflow) error {
173 e.l.Info("setting up workflow", "workflow", wid)
174
175 _, err := e.docker.NetworkCreate(ctx, networkName(wid), network.CreateOptions{
176 Driver: "bridge",
177 })
178 if err != nil {
179 return err
180 }
181 e.registerCleanup(wid, func(ctx context.Context) error {
182 err := e.docker.NetworkRemove(ctx, networkName(wid))
183 if err != nil {
184 return fmt.Errorf("removing network: %w", err)
185 }
186 return nil
187 })
188
189 addl := wf.Data.(addlFields)
190
191 reader, err := e.docker.ImagePull(ctx, addl.image, image.PullOptions{})
192 if err != nil {
193 e.l.Error("pipeline image pull failed!", "image", addl.image, "workflowId", wid, "error", err.Error())
194
195 return fmt.Errorf("pulling image: %w", err)
196 }
197 defer reader.Close()
198 io.Copy(os.Stdout, reader)
199
200 resp, err := e.docker.ContainerCreate(ctx, &container.Config{
201 Image: addl.image,
202 Cmd: []string{"cat"},
203 OpenStdin: true, // so cat stays alive :3
204 Tty: false,
205 Hostname: "spindle",
206 WorkingDir: workspaceDir,
207 Labels: map[string]string{
208 "sh.tangled.pipeline/workflow_id": wid.String(),
209 },
210 // TODO(winter): investigate whether environment variables passed here
211 // get propagated to ContainerExec processes
212 }, &container.HostConfig{
213 Mounts: []mount.Mount{
214 {
215 Type: mount.TypeTmpfs,
216 Target: "/tmp",
217 ReadOnly: false,
218 TmpfsOptions: &mount.TmpfsOptions{
219 Mode: 0o1777, // world-writeable sticky bit
220 Options: [][]string{
221 {"exec"},
222 },
223 },
224 },
225 },
226 ReadonlyRootfs: false,
227 CapDrop: []string{"ALL"},
228 CapAdd: []string{"CAP_DAC_OVERRIDE", "CAP_CHOWN", "CAP_FOWNER", "CAP_SETUID", "CAP_SETGID"},
229 SecurityOpt: []string{"no-new-privileges"},
230 ExtraHosts: []string{"host.docker.internal:host-gateway"},
231 }, nil, nil, "")
232 if err != nil {
233 return fmt.Errorf("creating container: %w", err)
234 }
235 e.registerCleanup(wid, func(ctx context.Context) error {
236 err := e.docker.ContainerStop(ctx, resp.ID, container.StopOptions{})
237 if err != nil {
238 return fmt.Errorf("stopping container: %w", err)
239 }
240
241 err = e.docker.ContainerRemove(ctx, resp.ID, container.RemoveOptions{
242 RemoveVolumes: true,
243 RemoveLinks: false,
244 Force: false,
245 })
246 if err != nil {
247 return fmt.Errorf("removing container: %w", err)
248 }
249 return nil
250 })
251
252 err = e.docker.ContainerStart(ctx, resp.ID, container.StartOptions{})
253 if err != nil {
254 return fmt.Errorf("starting container: %w", err)
255 }
256
257 mkExecResp, err := e.docker.ContainerExecCreate(ctx, resp.ID, container.ExecOptions{
258 Cmd: []string{"mkdir", "-p", workspaceDir, homeDir},
259 AttachStdout: true, // NOTE(winter): pretty sure this will make it so that when stdout read is done below, mkdir is done. maybe??
260 AttachStderr: true, // for good measure, backed up by docker/cli ("If -d is not set, attach to everything by default")
261 })
262 if err != nil {
263 return err
264 }
265
266 // This actually *starts* the command. Thanks, Docker!
267 execResp, err := e.docker.ContainerExecAttach(ctx, mkExecResp.ID, container.ExecAttachOptions{})
268 if err != nil {
269 return err
270 }
271 defer execResp.Close()
272
273 // This is apparently best way to wait for the command to complete.
274 _, err = io.ReadAll(execResp.Reader)
275 if err != nil {
276 return err
277 }
278
279 execInspectResp, err := e.docker.ContainerExecInspect(ctx, mkExecResp.ID)
280 if err != nil {
281 return err
282 }
283
284 if execInspectResp.ExitCode != 0 {
285 return fmt.Errorf("mkdir exited with exit code %d", execInspectResp.ExitCode)
286 } else if execInspectResp.Running {
287 return errors.New("mkdir is somehow still running??")
288 }
289
290 addl.container = resp.ID
291 wf.Data = addl
292
293 return nil
294}
295
296func (e *Engine) RunStep(ctx context.Context, wid models.WorkflowId, w *models.Workflow, idx int, secrets []secrets.UnlockedSecret, wfLogger *models.WorkflowLogger) error {
297 addl := w.Data.(addlFields)
298 workflowEnvs := ConstructEnvs(w.Environment)
299 // TODO(winter): should SetupWorkflow also have secret access?
300 // IMO yes, but probably worth thinking on.
301 for _, s := range secrets {
302 workflowEnvs.AddEnv(s.Key, s.Value)
303 }
304
305 step := w.Steps[idx]
306
307 select {
308 case <-ctx.Done():
309 return ctx.Err()
310 default:
311 }
312
313 envs := append(EnvVars(nil), workflowEnvs...)
314 if nixStep, ok := step.(Step); ok {
315 for k, v := range nixStep.environment {
316 envs.AddEnv(k, v)
317 }
318 }
319 envs.AddEnv("HOME", homeDir)
320
321 mkExecResp, err := e.docker.ContainerExecCreate(ctx, addl.container, container.ExecOptions{
322 Cmd: []string{"bash", "-c", step.Command()},
323 AttachStdout: true,
324 AttachStderr: true,
325 Env: envs,
326 })
327 if err != nil {
328 return fmt.Errorf("creating exec: %w", err)
329 }
330
331 // start tailing logs in background
332 tailDone := make(chan error, 1)
333 go func() {
334 tailDone <- e.tailStep(ctx, wfLogger, mkExecResp.ID, wid, idx, step)
335 }()
336
337 select {
338 case <-tailDone:
339
340 case <-ctx.Done():
341 // cleanup will be handled by DestroyWorkflow, since
342 // Docker doesn't provide an API to kill an exec run
343 // (sure, we could grab the PID and kill it ourselves,
344 // but that's wasted effort)
345 e.l.Warn("step timed out", "step", step.Name())
346
347 <-tailDone
348
349 return engine.ErrTimedOut
350 }
351
352 select {
353 case <-ctx.Done():
354 return ctx.Err()
355 default:
356 }
357
358 execInspectResp, err := e.docker.ContainerExecInspect(ctx, mkExecResp.ID)
359 if err != nil {
360 return err
361 }
362
363 if execInspectResp.ExitCode != 0 {
364 inspectResp, err := e.docker.ContainerInspect(ctx, addl.container)
365 if err != nil {
366 return err
367 }
368
369 e.l.Error("workflow failed!", "workflow_id", wid.String(), "exit_code", execInspectResp.ExitCode, "oom_killed", inspectResp.State.OOMKilled)
370
371 if inspectResp.State.OOMKilled {
372 return ErrOOMKilled
373 }
374 return engine.ErrWorkflowFailed
375 }
376
377 return nil
378}
379
380func (e *Engine) tailStep(ctx context.Context, wfLogger *models.WorkflowLogger, execID string, wid models.WorkflowId, stepIdx int, step models.Step) error {
381 if wfLogger == nil {
382 return nil
383 }
384
385 // This actually *starts* the command. Thanks, Docker!
386 logs, err := e.docker.ContainerExecAttach(ctx, execID, container.ExecAttachOptions{})
387 if err != nil {
388 return err
389 }
390 defer logs.Close()
391
392 _, err = stdcopy.StdCopy(
393 wfLogger.DataWriter(stepIdx, "stdout"),
394 wfLogger.DataWriter(stepIdx, "stderr"),
395 logs.Reader,
396 )
397 if err != nil && err != io.EOF && !errors.Is(err, context.DeadlineExceeded) {
398 return fmt.Errorf("failed to copy logs: %w", err)
399 }
400
401 return nil
402}
403
404func (e *Engine) DestroyWorkflow(ctx context.Context, wid models.WorkflowId) error {
405 fns := e.drainCleanups(wid)
406
407 for _, fn := range fns {
408 if err := fn(ctx); err != nil {
409 e.l.Error("failed to cleanup workflow resource", "workflowId", wid, "error", err)
410 }
411 }
412 return nil
413}
414
415func (e *Engine) registerCleanup(wid models.WorkflowId, fn cleanupFunc) {
416 e.cleanupMu.Lock()
417 defer e.cleanupMu.Unlock()
418
419 key := wid.String()
420 e.cleanup[key] = append(e.cleanup[key], fn)
421}
422
423func (e *Engine) drainCleanups(wid models.WorkflowId) []cleanupFunc {
424 e.cleanupMu.Lock()
425 key := wid.String()
426
427 fns := e.cleanup[key]
428 delete(e.cleanup, key)
429 e.cleanupMu.Unlock()
430
431 return fns
432}
433
434func networkName(wid models.WorkflowId) string {
435 return fmt.Sprintf("workflow-network-%s", wid)
436}